2021SC@SDUSC
前言
本文分析yolov5代码中torch_utils.py文件,主要是一些封装好的工具类。
torch_distributed_zero_frist函数
@contextmanager
def torch_distributed_zero_first(local_rank: int):
"""
Decorator to make all processes in distributed training wait for each local_master to do something.
"""
if local_rank not in [-1, 0]:
dist.barrier(device_ids=[local_rank])
yield
if local_rank == 0:
dist.barrier(device_ids=[0])
该函数用于处理模型分布式训练时同步问题,如果当前进程不是主进程,设置一个阻塞栅栏,让此进程处于等待状态,等待所有进程到达栅栏处;当所有进程都到达栅栏处后释放栅栏。
date_modified函数
def date_modified(path=__file__):
# return human-readable file modification date, i.e. '2021-3-26'
t = datetime.datetime.fromtimestamp(Path(path).stat().st_mtime)
return f'{t.year}-{t.month}-{t.day}'
将系统时间改为人类可读的时间格式
select_device函数
def select_device(device='', batch_size=None):
# device = 'cpu' or '0' or '0,1,2,3'
s = f'YOLOv5 🚀 {git_describe() or date_modified()} torch {torch.__version__} ' # string
device = str(device).strip().lower().replace('cuda:', '') # to string, 'cuda:0' to '0'
cpu = device == 'cpu'
if cpu:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1' # force torch.cuda.is_available() = False
elif device: # non-cpu device requested
os.environ['CUDA_VISIBLE_DEVICES'] = device # set environment variable
assert torch.cuda.is_available(), f'CUDA unavailable, invalid device {device} requested' # check availability
cuda = not cpu and torch.cuda.is_available()
if cuda:
devices = device.split(',') if device else '0' # range(torch.cuda.device_count()) # i.e. 0,1,6,7
n = len(devices) # device count
if n > 1 and batch_size: # check batch_size is divisible by device_count
assert batch_size % n == 0, f'batch-size {batch_size} not multiple of GPU count {n}'
space = ' ' * (len(s) + 1)
for i, d in enumerate(devices):
p = torch.cuda.get_device_properties(i)
s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / 1024 ** 2}MB)\n" # bytes to MB
else:
s += 'CPU\n'
LOGGER.info(s.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else s) # emoji-safe
return torch.device('cuda:0' if cuda else 'cpu')
该函数实现选择训练的设备,当参数device不为空时,选择指定的设备,即cpu或gou;当参数device为空时根据机器情况自动选择设备,优先选择gpu作为训练设备。
time_sync函数
def time_sync():
# pytorch-accurate time
if torch.cuda.is_available():
torch.cuda.synchronize()
return time.time()
该函数用于在进行分布式操作时精确地获取当前时间。
profile函数
def profile(input, ops, n=10, device=None):
# YOLOv5 speed/memory/FLOPs profiler
#
# Usage:
# input = torch.randn(16, 3, 640, 640)
# m1 = lambda x: x * torch.sigmoid(x)
# m2 = nn.SiLU()
# profile(input, [m1, m2], n=100) # profile over 100 iterations
results = []
logging.basicConfig(format="%(message)s", level=logging.INFO)
device = device or select_device()
print(f"{'Params':>12s}{'GFLOPs':>12s}{'GPU_mem (GB)':>14s}{'forward (ms)':>14s}{'backward (ms)':>14s}"
f"{'input':>24s}{'output':>24s}")
for x in input if isinstance(input, list) else [input]:
x = x.to(device)
x.requires_grad = True
for m in ops if isinstance(ops, list) else [ops]:
m = m.to(device) if hasattr(m, 'to') else m # device
m = m.half() if hasattr(m, 'half') and isinstance(x, torch.Tensor) and x.dtype is torch.float16 else m
tf, tb, t = 0., 0., [0., 0., 0.] # dt forward, backward
try:
flops = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2 # GFLOPs
except:
flops = 0
try:
for _ in range(n):
t[0] = time_sync()
y = m(x)
t[1] = time_sync()
try:
_ = (sum([yi.sum() for yi in y]) if isinstance(y, list) else y).sum().backward()
t[2] = time_sync()
except Exception as e: # no backward method
print(e)
t[2] = float('nan')
tf += (t[1] - t[0]) * 1000 / n # ms per op forward
tb += (t[2] - t[1]) * 1000 / n # ms per op backward
mem = torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0 # (GB)
s_in = tuple(x.shape) if isinstance(x, torch.Tensor) else 'list'
s_out = tuple(y.shape) if isinstance(y, torch.Tensor) else 'list'
p = sum(list(x.numel() for x in m.parameters())) if isinstance(m, nn.Module) else 0 # parameters
print(f'{p:12}{flops:12.4g}{mem:>14.3f}{tf:14.4g}{tb:14.4g}{str(s_in):>24s}{str(s_out):>24s}')
results.append([p, flops, mem, tf, tb, s_in, s_out])
except Exception as e:
print(e)
results.append(None)
torch.cuda.empty_cache()
return results
该函数用于输出训练过程中的一些相关信息,如前向传播时间、反向传播时间、输入变量的shape、输出变量的shape等。
initialize_weights函数
def initialize_weights(model):
for m in model.modules():
t = type(m)
if t is nn.Conv2d:
pass # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif t is nn.BatchNorm2d:
m.eps = 1e-3
m.momentum = 0.03
elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6]:
m.inplace = True
该函数实现训练参数的初始化,BN的eps设为1e-3,momentum设为0.03,另外激活函数inplace都为True
find_modules函数
def find_modules(model, mclass=nn.Conv2d):
# Finds layer indices matching module class 'mclass'
return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)]
该函数实现在整个模型中返回某一个模块的索引值
fuse_conv_and_bn函数
def fuse_conv_and_bn(conv, bn):
# Fuse convolution and batchnorm layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/
fusedconv = nn.Conv2d(conv.in_channels,
conv.out_channels,
kernel_size=conv.kernel_size,
stride=conv.stride,
padding=conv.padding,
groups=conv.groups,
bias=True).requires_grad_(False).to(conv.weight.device)
# prepare filters
w_conv = conv.weight.clone().view(conv.out_channels, -1)
w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape))
# prepare spatial bias
b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias
b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps))
fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn)
return fusedconv
该函数在推理测试阶段,将卷积层和BN层进行融合,加速推理。
model_info函数
def model_info(model, verbose=False, img_size=640):
# Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320]
n_p = sum(x.numel() for x in model.parameters()) # number parameters
n_g = sum(x.numel() for x in model.parameters() if x.requires_grad) # number gradients
if verbose:
print('%5s %40s %9s %12s %20s %10s %10s' % ('layer', 'name', 'gradient', 'parameters', 'shape', 'mu', 'sigma'))
for i, (name, p) in enumerate(model.named_parameters()):
name = name.replace('module_list.', '')
print('%5g %40s %9s %12g %20s %10.3g %10.3g' %
(i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std()))
try: # FLOPs
from thop import profile
stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32
img = torch.zeros((1, model.yaml.get('ch', 3), stride, stride), device=next(model.parameters()).device) # input
flops = profile(deepcopy(model), inputs=(img,), verbose=False)[0] / 1E9 * 2 # stride GFLOPs
img_size = img_size if isinstance(img_size, list) else [img_size, img_size] # expand if int/float
fs = ', %.1f GFLOPs' % (flops * img_size[0] / stride * img_size[1] / stride) # 640x640 GFLOPs
except (ImportError, Exception):
fs = ''
LOGGER.info(f"Model Summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}")
该函数用来输出模型的所有信息,包括所有层数量,模型总参数量,需要求梯度的总参数量,img_size大小等。
load_classifier函数
def load_classifier(name='resnet101', n=2):
# Loads a pretrained model reshaped to n-class output
model = torchvision.models.__dict__[name](pretrained=True)
# ResNet model properties
# input_size = [3, 224, 224]
# input_space = 'RGB'
# input_range = [0, 1]
# mean = [0.485, 0.456, 0.406]
# std = [0.229, 0.224, 0.225]
# Reshape output to n classes
filters = model.fc.weight.shape[1]
model.fc.bias = nn.Parameter(torch.zeros(n), requires_grad=True)
model.fc.weight = nn.Parameter(torch.zeros(n, filters), requires_grad=True)
model.fc.out_features = n
return model
该函数用于在检测后进行二次分类,将resnet的fc层的输出改为n,即需要分类的类别数。
scale_img函数
def scale_img(img, ratio=1.0, same_shape=False, gs=32): # img(16,3,256,416)
# scales img(bs,3,y,x) by ratio constrained to gs-multiple
if ratio == 1.0:
return img
else:
h, w = img.shape[2:]
s = (int(h * ratio), int(w * ratio)) # new size
img = F.interpolate(img, size=s, mode='bilinear', align_corners=False) # resize
if not same_shape: # pad/crop img
h, w = [math.ceil(x * ratio / gs) * gs for x in (h, w)]
return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447) # value = imagenet mean
该函数用于对图片进行缩放,ratio为需要进行缩放的比例,当same_shape为true时要求缩放后的图片大小是gs的倍数。
EarlyStopping类
class EarlyStopping:
# YOLOv5 simple early stopper
def __init__(self, patience=30):
self.best_fitness = 0.0 # i.e. mAP
self.best_epoch = 0
self.patience = patience or float('inf') # epochs to wait after fitness stops improving to stop
self.possible_stop = False # possible stop may occur next epoch
def __call__(self, epoch, fitness):
if fitness >= self.best_fitness: # >= 0 to allow for early zero-fitness stage of training
self.best_epoch = epoch
self.best_fitness = fitness
delta = epoch - self.best_epoch # epochs without improvement
self.possible_stop = delta >= (self.patience - 1) # possible stop may occur next epoch
stop = delta >= self.patience # stop training if patience exceeded
if stop:
LOGGER.info(f'EarlyStopping patience {self.patience} exceeded, stopping training.')
return stop
该类是用于在训练时早停策略的,也就是如果当前的评估指标比前几个epoch的值还要差的话就停止训练。patience为容忍度,也就是最好的epoch与当前epoch之差的最大值,当当前epochs与最好的epoch之差大于这个值,并且当前epoch的指标不是最好的,那么久停止训练。早停策略是防止训练过拟合的一个常用的方法。
总结
这一篇文章分析了一些使用的pytorch工具方法,包括了一些训练以及测试时的策略,都是与pytorch有关的。